Biblio

Minimal adversarial perturbations added to inputs have been shown to be effective at fooling deep neural networks. In this paper, we introduce several innovations that make white-box targeted attacks follow the intuition of the attacker's goal: to trick the model to assign a higher probability to the target class than to any other, while staying within a specified distance from the original input. First, we propose a new loss function that explicitly captures the goal of targeted attacks, in particular, by using the logits of all classes instead of just a subset, as is common. We show that Auto-PGD with this loss function finds more adversarial examples than it does with other commonly used loss functions. Second, we propose a new attack method that uses a further developed version of our loss function capturing both the misclassification objective and the L∞ distance limit ϵ. This new attack method is relatively 1.5--4.2% more successful on the CIFAR10 dataset and relatively 8.2--14.9% more successful on the ImageNet dataset, than the next best state-of-the-art attack. We confirm using statistical tests that our attack outperforms state-of-the-art attacks on different datasets and values of ϵ and against different defenses.

Programming language design requires making many usability-related design decisions. However, existing HCI methods can be impractical to apply to programming languages: they have high iteration costs, programmers require significant learning time, and user performance has high variance. To address these problems, we adapted both formative and summative HCI methods to make them more suitable for programming language design. We integrated these methods into a new process, PLIERS, for designing programming languages in a user-centered way. We evaluated PLIERS by using it to design two new programming languages. Glacier extends Java to enable programmers to express immutability properties effectively and easily. Obsidian is a language for blockchains that includes verification of critical safety properties. Summative usability studies showed that programmers were able to program effectively in both languages after short training periods.

For realistic self-adaptive systems, multiple quality attributes need to be considered and traded off against each other. These quality attributes are commonly encoded in a utility function, for instance, a weighted sum of relevant objectives. [Question/problem:] The research agenda for requirements engineering for self-adaptive systems has raised the need for decision-making techniques that consider the trade-offs and priorities of multiple objectives. Human stakeholders need to be engaged in the decision-making process so that the relative importance of each objective can be correctly elicited. [Principal ideas/results:] This research preview paper presents a method that supports multiple stakeholders in prioritizing relevant quality attributes, negotiating priorities to reach an agreement, and giving input to define utility functions for self-adaptive systems. [Contribution:] The proposed method constitutes a lightweight solution for utility function definition. It can be applied by practitioners and researchers who aim to develop self-adaptive systems that meet stakeholders’ requirements. We present details of our plan to study the application of our method using a case study.

We propose Black-Box IoT (BBox-IoT), a new ultra-lightweight black-box system for authenticating and storing IoT data. BBox-IoT is tailored for deployment on IoT devices (including low-Size Weight and Power sensors) which are extremely constrained in terms of computation, storage, and power. By utilizing core Blockchain principles, we ensure that the collected data is immutable and tamper-proof while preserving data provenance and non-repudiation. To realize BBox-IoT, we designed and implemented a novel chain-based hash signature scheme which only requires hashing operations and removes all synchronicity dependencies between signer and verifier. Our approach enables low-SWaP devices to authenticate removing reliance on clock synchronization. Our evaluation results show that BBox-IoT is practical in Industrial Internet of Things (IIoT) environments: even devices equipped with 16MHz microcontrollers and 2KB memory can broadcast their collected data without requiring heavy cryptographic operations or synchronicity assumptions. Finally, when compared to industry standard ECDSA, our approach is two and three orders of magnitude faster for signing and verification operations respectively. Thus, we are able to increase the total number of signing operations by more than 5000% for the same amount of power.

Many self-adaptive systems benefit from human involvement, where human operators can complement the capabilities of systems (e.g., by supervising decisions, or performing adaptations and tasks involving physical changes that cannot be automated). However, insufficient preparation (e.g., lack of task context comprehension) may hinder the effectiveness of human involvement, especially when operators are unexpectedly interrupted to perform a new task. Preparatory notification of a task provided in advance can sometimes help human operators focus their attention on the forthcoming task and understand its context before task execution, hence improving effectiveness. Nevertheless, deciding when to use preparatory notification as a tactic is not obvious and entails considering different factors that include uncertainties induced by human operator behavior (who might ignore the notice message), human attributes (e.g., operator training level), and other information that refers to the state of the system and its environment. In this paper, informed by work in cognitive science on human attention and context management, we introduce a formal framework to reason about the usage of preparatory notifications in self-adaptive systems involving human operators. Our framework characterizes the effects of managing attention via task notification in terms of task context comprehension. We also build on our framework to develop an automated probabilistic reasoning technique able to determine when and in what form a preparatory notification tactic should be used to optimize system goals. We illustrate our approach in a representative scenario of human-robot collaborative goods delivery.

Two established approaches to engineer adaptive systems are architecture-based adaptation that uses a Monitor-Analysis-Planning-Executing (MAPE) loop that reasons over architectural models (aka Knowledge) to make adaptation decisions, and control-based adaptation that relies on principles of control theory (CT) to realize adaptation. Recently, we also observe a rapidly growing interest in applying machine learning (ML) to support different adaptation mechanisms. While MAPE and CT have particular characteristics and strengths to be applied independently, in this paper, we are concerned with the question of how these approaches are related with one another and whether combining them and supporting them with ML can produce better adaptive systems. We motivate the combined use of different adaptation approaches using a scenario of a cloud-based enterprise system and illustrate the analysis when combining the different approaches. To conclude, we offer a set of open questions for further research in this interesting area.

The question of whether “handling unanticipated changes is the ultimate challenge for self-adaptation” is impossible to evaluate without looking closely at what “unanticipated” means. In this position paper I try to bring a little clarity to this issue by arguing that the common distinction between “known unknowns” and “unknown unknowns” is too crude: for most systems there are changes that are not directly handled by “first-order” adaptation, but can, with appropriate engineering, be addressed naturally through “second-order” adaptation. I explain what I mean by this and consider ways in which such systems might be engineered.

Motivated by the transformative impact of deep neural networks (DNNs) in various domains, researchers and anti-virus vendors have proposed DNNs for malware detection from raw bytes that do not require manual feature engineering. In this work, we propose an attack that interweaves binary-diversification techniques and optimization frameworks to mislead such DNNs while preserving the functionality of binaries. Unlike prior attacks, ours manipulates instructions that are a functional part of the binary, which makes it particularly challenging to defend against. We evaluated our attack against three DNNs in white- and black-box settings, and found that it often achieved success rates near 100%. Moreover, we found that our attack can fool some commercial anti-viruses, in certain cases with a success rate of 85%. We explored several defenses, both new and old, and identified some that can foil over 80% of our evasion attempts. However, these defenses may still be susceptible to evasion by attacks, and so we advocate for augmenting malware-detection systems with methods that do not rely on machine learning.

Alloy is a declarative modeling language based on a first-order relational logic. Its constraint-based analysis has enabled a wide range of applications in software engineering, including configuration synthesis, bug finding, test-case generation, and security analysis. Certain types of analysis tasks in these domains involve finding an optimal solution. For example, in a network configuration problem, instead of finding any valid configuration, it may be desirable to find one that is most permissive (i.e., it permits a maximum number of packets). Due to its dependence on SAT, however, Alloy cannot be used to specify and analyze these types of problems. We propose AlloyMax, an extension of Alloy with a capability to express and analyze problems with optimal solutions. AlloyMax introduces (1) a small addition of language constructs that can be used to specify a wide range of problems that involve optimality and (2) a new analysis engine that leverages a Maximum Satisfiability (MaxSAT) solver to generate optimal solutions. To enable this new type of analysis, we show how a specification in a first-order relational logic can be translated into an input format of MaxSAT solvers—namely, a Boolean formula in weighted conjunctive normal form (WCNF). We demonstrate the applicability and scalability of AlloyMax on a benchmark of problems. To our knowledge, AlloyMax is the first approach to enable analysis with optimality in a relational modeling language, and we believe that AlloyMax has the potential to bring a wide range of new applications to Alloy.

In software design, guaranteeing the correctness of run-time system behavior while achieving an acceptable balance among multiple quality attributes remains a challenging problem. Moreover, providing guarantees about the satisfaction of those requirements when systems are subject to uncertain environments is even more challenging. While recent developments in architectural analysis techniques can assist architects in exploring the satisfaction of quantitative guarantees across the design space, existing approaches are still limited because they do not explicitly link design decisions to satisfaction of quality requirements. Furthermore, the amount of information they yield can be overwhelming to a human designer, making it difficult to distinguish the forest through the trees. In this paper, we present an approach to analyzing architectural design spaces that addresses these limitations and provides a basis to enable the explainability of design tradeoffs. Our approach combines dimensionality reduction techniques employed in machine learning pipelines with quantitative verification to enable architects to understand how design decisions contribute to the satisfaction of strict quantitative guarantees under uncertainty across the design space. Our results show feasibility of the approach in two case studies and evidence that dimensionality reduction is a viable approach to facilitate comprehension of tradeoffs in poorly-understood design spaces.

Explainability refers to the degree to which a software system’s actions or solutions can be understood by humans. Giving humans the right amount of explanation at the right time is an important factor in maximizing the effective collaboration between an adaptive system and humans during interaction. However, explanations come with costs, such as the required time of explanation and humans’ response time. Hence it is not always clear whether explanations will improve overall system utility and, if so, how the system should effectively provide explanation to humans, particularly given that different humans may benefit from different amounts and frequency of explanation. To provide a partial basis for making such decisions, this paper defines a formal framework that incorporates human personality traits as one of the important elements in guiding automated decision- making about the proper amount of explanation that should be given to the human to improve the overall system utility. Specifically, we use probabilistic model analysis to determine how to utilize explanations in an effective way. To illustrate our approach, Grid – a virtual human and system interaction game -- is developed to represent scenarios for human-systems collaboration and to demonstrate how a human’s personality traits can be used as a factor to consider for systems in providing appropriate explanations.

Today’s world is witnessing a shift from human-written software to machine-learned software, with the rise of systems that rely on machine learning. These systems typically operate in non-static environments, which are prone to unexpected changes, as is the case of self-driving cars and enterprise systems. In this context, machine-learned software can misbehave. Thus, it is paramount that these systems are capable of detecting problems with their machined-learned components and adapt themselves to maintain desired qualities. For instance, a fraud detection system that cannot adapt its machine-learned model to efficiently cope with emerging fraud patterns or changes in the volume of transactions is subject to losses of millions of dollars. In this paper, we take a first step towards the development of a framework aimed to self-adapt systems that rely on machine-learned components. We describe: (i) a set of causes of machine-learned component misbehavior and a set of adaptation tactics inspired by the literature on machine learning, motivating them with the aid of a running example; (ii) the required changes to the MAPE-K loop, a popular control loop for self-adaptive systems; and (iii) the challenges associated with developing this framework. We conclude the paper with a set of research questions to guide future work.

Cyber-Physical Systems (CPS) integrate computational and physical components. With the digitisation of society and industry and the progressing integration of systems, CPS need to become “smarter” in the sense that they can adapt and learn to handle new and unexpected conditions, and improve over time. Smarter CPS present a combination of challenges that existing engineering methods have difficulties addressing: intertwined digital, physical and social spaces, need for heterogeneous modelling formalisms, demand for context-tied cooperation to achieve system goals, widespread uncertainty and disruptions in changing contexts, inherent human constituents, and continuous encounter with new situations. While approaches have been put forward to deal with some of these challenges, a coherent perspective on engineering smarter CPS is lacking. In this paper, we present six engineering principles for addressing the challenges of smarter CPS. As smarter CPS are software-intensive systems, we approach them from a software engineering perspective with the angle of self-adaptation that offers an effective approach to deal with run-time change. The six principles create an integrated landscape for the engineering and operation of smarter CPS.

Self-adaptive systems overcome many of the limitations of human supervision in complex software-intensive systems by endowing them with the ability to automatically adapt their structure and behavior in the presence of runtime changes. However, adaptation in some classes of systems (e.g., safetycritical) can benefit by receiving information from humans (e.g., acting as sophisticated sensors, decision-makers), or by involving them as system-level effectors to execute adaptations (e.g., when automation is not possible, or as a fallback mechanism). However, human participants are influenced by factors external to the system (e.g., training level, fatigue) that affect the likelihood of success when they perform a task, its duration, or even if they are willing to perform it in the first place. Without careful consideration of these factors, it is unclear how to decide when to involve humans in adaptation, and in which way. In this paper, we investigate how the explicit modeling of human participants can provide a better insight into the trade-offs of involving humans in adaptation. We contribute a formal framework to reason about human involvement in self-adaptation, focusing on the role of human participants as actors (i.e., effectors) during the execution stage of adaptation. The approach consists of: (i) a language to express adaptation models that capture factors affecting human behavior and its interactions with the system, and (ii) a formalization of these adaptation models as stochastic multiplayer games (SMGs) that can be used to analyze humansystem-environment interactions. We illustrate our approach in an adaptive industrial middleware used to monitor and manage sensor networks in renewable energy production plants.

To protect against misuse of passwords compromised in a breach, consumers should promptly change affected passwords and any similar passwords on other accounts. Ideally, affected companies should strongly encourage this behavior and have mechanisms in place to mitigate harm. In order to make recommendations to companies about how to help their users perform these and other security-enhancing actions after breaches, we must first have some understanding of the current effectiveness of companies’ post-breach practices. To study the effectiveness of password-related breach notifications and practices enforced after a breach, we examine—based on real-world password data from 249 participants—whether and how constructively participants changed their passwords after a breach announcement. Of the 249 participants, 63 had accounts on breached domains; only 33% of the 63 changed their passwords and only 13% (of 63) did so within three months of the announcement. New passwords were on average 1.3× stronger than old passwords (when comparing log10-transformed strength), though most were weaker or of equal strength. Concerningly, new passwords were overall more similar to participants’ other passwords, and participants rarely changed passwords on other sites even when these were the same or similar to their password on the breached domain. Our results highlight the need for more rigorous passwordchanging requirements following a breach and more effective breach notifications that deliver comprehensive advice.

Many self-adaptive systems benefit from human involvement and oversight, where a human operator can provide expertise not available to the system and can detect problems that the system is unaware of. One way of achieving this is by placing the human operator on the loop – i.e., providing supervisory oversight and intervening in the case of questionable adaptation decisions. To make such interaction effective, explanation is sometimes helpful to allow the human to understand why the system is making certain decisions and calibrate confidence from the human perspective. However, explanations come with costs in terms of delayed actions and the possibility that a human may make a bad judgement. Hence, it is not always obvious whether explanations will improve overall utility and, if so, what kinds of explanation to provide to the operator. In this work, we define a formal framework for reasoning about explanations of adaptive system behaviors and the conditions under which they are warranted. Specifically, we characterize explanations in terms of explanation content, effect, and cost. We then present a dynamic adaptation approach that leverages a probabilistic reasoning technique to determine when the explanation should be used in order to improve overall system utility.

In this work we develop a novel Bayesian neural network methodology to achieve strong adversarial robustness without the need for online adversarial training. Unlike previous efforts in this direction, we do not rely solely on the stochasticity of network weights by minimizing the divergence between the learned parameter distribution and a prior. Instead, we additionally require that the model maintain some expected uncertainty with respect to all input covariates. We demonstrate that by encouraging the network to distribute evenly across inputs, the network becomes less susceptible to localized, brittle features which imparts a natural robustness to targeted perturbations. We show empirical robustness on several benchmark datasets.

End-users’ trust in automated agents is important as automated decision-making and planning is increasingly used in many aspects of people’s lives. In real-world applications of planning, multiple optimization objectives are often involved. Thus, planning agents’ decisions can involve complex tradeoffs among competing objectives. It can be difficult for the end-users to understand why an agent decides on a particular planning solution on the basis of its objective values. As a result, the users may not know whether the agent is making the right decisions, and may lack trust in it. In this work, we contribute an approach, based on contrastive explanation, that enables a multi-objective MDP planning agent to explain its decisions in a way that communicates its tradeoff rationale in terms of the domain-level concepts. We conduct a human subjects experiment to evaluate the effectiveness of our explanation approach in a mobile robot navigation domain. The results show that our approach significantly improves the users’ understanding, and confidence in their understanding, of the tradeoff rationale of the planning agent.

Programming languages exist to enable people to create and maintain software as effectively as possible. They are subject to two very different sets of requirements: first, the need to provide strong safety guarantees to protect against bugs; and second, the need for users to effectively and efficiently write software that meets their functional and quality requirements. This thesis argues that fusing formal methods for reasoning about programming languages with user-centered design methods is a practical approach to designing languages that make programmers more effective. By doing so, designers can create safer languages that are more effective for programmers than existing languages. The thesis is substantiated by the introduction of PLIERS: Programming Language Iterative Evaluation and Refinement System. PLIERS is a process for designing programming languages that integrates formal methods with user-centered design. The hypothesis that PLIERS is beneficial is supported by two language design projects. In these projects, I show how PLIERS benefits the programming language design process. Glacier is an extension to Java that enforces transitive class immutability, which is a much stronger property than that provided by languages that are in use today. Although there are hundreds of possible ways of restricting changes to state in programming languages, Glacier occupies a point in the design space that was justified by interview studies with professional software engineers. I evaluated Glacier in case studies, showing that it is expressive enough for some real-world applications. I also evaluated Glacier in lab studies and compared it to Java’s final keyword, finding both that Glacier is more effective at expressing immutability than final and that Glacier detects bugs that users are likely to insert in code. Blockchains are a distributed computing platform that aim to enable safe computation among users who do not necessarily trust each other. To improve safety relative to existing languages, in which programmers have repeatedly deployed software with serious bugs, I designed Obsidian, a new programming language for blockchain application development. From observations about typical blockchain applications, I derived two features that motivated the design of Obsidian. First, blockchain applications typically implement state machines, which support different operations in different states. Obsidian uses typestate, which lifts dynamic state into static types, to provide static guarantees regarding object state. Second, blockchain applications frequently manipulate resources, such as virtual currency. Obsidian provides a notion of ownership, which distinguishes one reference from all others. Obsidian supports resources via linear types, which ensure that owning references to resources are not accidentally lost. The combination of resources and typestate results in a novel set of design problems for the type system; although each idea has been explored individually, combining them requires unifying different perspectives on linearity. Furthermore, no language with either one of these features has been designed in a user-centered way or evaluated with users. Typical typestate systems have a complex set of permissions that provides safety properties, but these systems focused on expressiveness rather than on usability. Obsidian integrates typestate and resources in a novel way, resulting in a new type system design with a focus on simplicity and usability while retaining the desired safety properties. Obsidian is based on a core calculus I designed, Silica, for which I proved type soundness. In order to make Obsidian as usable as possible, I based its design on the results of formative studies with programmers. I evaluated Obsidian with two case studies, showing that Obsidian can be used to implement relevant programs. I also conducted a randomized controlled trial comparing Obsidian to Solidity, a popular language for writing smart contracts. I found that most of the Obsidian participants learned Obsidian and completed programming tasks after only a short training period; further, in one task, 70% of the participants who used Solidity accidentally inserted bugs that Obsidian’s compiler would have detected. Finally, Obsidian participants completed significantly more tasks correctly than did Solidity participants.

Some blockchain programs (smart contracts) have included serious security vulnerabilities. Obsidian is a new typestate-oriented programming language that uses a strong type system to rule out some of these vulnerabilities. Although Obsidian was designed to promote usability to make it as easy as possible to write programs, strong type systems can cause a language to be difficult to use. In particular, ownership, typestate, and assets, which Obsidian uses to provide safety guarantees, have not seen broad adoption together in popular languages and result in significant usability challenges. We performed an empirical study with 20 participants comparing Obsidian to Solidity, which is the language most commonly used for writing smart contracts today. We observed that Obsidian participants were able to successfully complete more of the programming tasks than the Solidity participants. We also found that the Solidity participants commonly inserted asset-related bugs, which Obsidian detects at compile time.

Blockchain platforms are coming into use for processing critical transactions among participants who have not established mutual trust. Many blockchains are programmable, supporting smart contracts, which maintain persistent state and support transactions that transform the state. Unfortunately, bugs in many smart contracts have been exploited by hackers. Obsidian is a novel programming language with a type system that enables static detection of bugs that are common in smart contracts today. Obsidian is based on a core calculus, Silica, for which we proved type soundness. Obsidian uses typestate to detect improper state manipulation and uses linear types to detect abuse of assets. We integrated a permissions system that encodes a notion of ownership to allow for safe, flexible aliasing. We describe two case studies that evaluate Obsidian’s applicability to the domains of parametric insurance and supply chain management, finding that Obsidian’s type system facilitates reasoning about high-level states and ownership of resources. We compared our Obsidian implementation to a Solidity implementation, observing that the Solidity implementation requires much boilerplate checking and tracking of state, whereas Obsidian does this work statically.

Blockchains host smart contracts for voting, tokens, and other purposes. Vulnerabilities in contracts are common, often leading to the loss of money. Psamathe is a new language we are designing around a new flow abstraction, reducing asset bugs and making contracts more concise than in existing languages. We present an overview of Psamathe, and discuss two example contracts in Psamathe and Solidity.

This paper proposes a new defense called $n$-ML against adversarial examples, i.e., inputs crafted by perturbing benign inputs by small amounts to induce misclassifications by classifiers. Inspired by $n$-version programming, $n$-ML trains an ensemble of $n$ classifiers, and inputs are classified by a vote of the classifiers in the ensemble. Unlike prior such approaches, however, the classifiers in the ensemble are trained specifically to classify adversarial examples differently, rendering it very difficult for an adversarial example to obtain enough votes to be misclassified. We show that $n$-ML roughly retains the benign classification accuracies of state-of-the-art models on the MNIST, CIFAR10, and GTSRB datasets, while simultaneously defending against adversarial examples with better resilience than the best defenses known to date and, in most cases, with lower classification-time overhead.

Motivated by the transformative impact of deep neural networks (DNNs) on different areas (e.g., image and speech recognition), researchers and anti-virus vendors are proposing end-to-end DNNs for malware detection from raw bytes that do not require manual feature engineering. Given the security sensitivity of the task that these DNNs aim to solve, it is important to assess their susceptibility to evasion.
In this work, we propose an attack that guides binary-diversification tools via optimization to mislead DNNs for malware detection while preserving the functionality of binaries. Unlike previous attacks on such DNNs, ours manipulates instructions that are a functional part of the binary, which makes it particularly challenging to defend against. We evaluated our attack against three DNNs in white-box and black-box settings, and found that it can often achieve success rates near 100%. Moreover, we found that our attack can fool some commercial anti-viruses, in certain cases with a success rate of 85%. We explored several defenses, both new and old, and identified some that can successfully prevent over 80% of our evasion attempts. However, these defenses may still be susceptible to evasion by adaptive attackers, and so we advocate for augmenting malware-detection systems with methods that do not rely on machine learning.

As Cyber-Physical Systems (CPSs) become more autonomous, it becomes harder for humans who interact with the CPSs to understand the behavior of the systems. Particularly for CPSs that must perform tasks while optimizing for multiple quality objectives and acting under uncertainty, it can be difficult for humans to understand the system behavior generated by an automated planner. This work-in-progress presents an approach at clarifying system behavior through interactive explanation by allowing end-users to ask Why and Why-Not questions about specific behaviors of the system, and providing answers in the form of contrastive explanation.